Subscribe to Dr. Granville's Weekly Digest

Based on my opinion. You are welcome to discuss. Note that most of these techniques have evolved over time (in the last 10 years) to the point where most drawbacks have been eliminated - making the updated tool far different and better than its original version. Typically, these bad techniques are still widely used.

  1. Linear regression. Relies on the normal, heteroscedasticity and other assumptions, does not capture highly non-linear, chaotic patterns. Prone to over-fitting. Parameters difficult to interpret. Very unstable when independent variables are highly correlated. Fixes: variable reduction, apply a transformation to your variables, use constrained regression (e.g. ridge or Lasso regression)
  2. Traditional decision trees. Very large decision trees are very unstable and impossible to interpret, and prone to over-fitting. Fix: combine multiple small decision trees together instead of using a large decision tree.
  3. Linear discriminant analysis. Used for supervised clustering. Bad technique because it assumes that clusters do not overlap, and are well separated by hyper-planes. In practice, they never do. Use density estimation techniques instead.
  4. K-means clustering. Used for clustering, tends to produce circular clusters. Does not work well with data points that are not a mixture of Gaussian distributions. 
  5. Neural networks. Difficult to interpret, unstable, subject to over-fitting.
  6. Maximum Likelihood estimation. Requires your data to fit with a prespecified probabilistic distribution. Not data-driven. In many cases the pre-specified Gaussian distribution is a terrible fit for your data.
  7. Density estimation in high dimensions. Subject to what is referred to as the curse of dimensionality. Fix: use (non parametric) kernel density estimators with adaptive bandwidths.
  8. Naive Bayes. Used e.g. in fraud and spam detection, and for scoring. Assumes that variables are independent, if not it will fail miserably. In the context of fraud or spam detection, variables (sometimes called rules) are highly correlated. Fix: group variables into independent clusters of variables (in each cluster, variables are highly correlated). Apply naive Bayes to the clusters. Or use data reduction techniques. Bad text mining techniques (e.g. basic "word" rules in spam detection) combined with naive Bayes produces absolutely terrible results with many false positives and false negatives.

And remember to use sound cross-validations techniques when testing models!

Additional comments:

The reasons why such poor models are still widely used are:

  1. Many University curricula still use outdated textbooks, thus many students are not exposed to better data science techniques
  2. People using black-box statistical software, not knowing the limitations, drawbacks, or how to correctly fine-tune the parameters and optimize the various knobs, or not understanding what the software actually produces.
  3. Government forcing regulated industries (pharmaceutical, banking, Basel) to use the same 30-year old SAS procedures for statistical compliance. For instance, better scoring methods for credit scoring, even if available in SAS, are not allowed and arbitrarily rejected by authorities. The same goes with clinical trials analyses submitted to the FDA, SAS being the mandatory software to be used for compliance, allowing the FDA to replicate analyses and results from pharmaceutical companies.
  4. Modern data sets are considerably more complex and different than the old data sets used when these techniques were initially developed. In short, these techniques have not been developed for modern data sets.
  5. There's no perfect statistical technique that would apply to all data sets, but there are many poor techniques.

In addition, poor cross-validation allows bad models to make the cut, by over-estimating the true lift to be expected in future data, the true accuracy or the true ROI outside the training set.  Good cross validations consist in

  • splitting your training set into multiple subsets (test and control subsets), 
  • include different types of clients and more recent data in the control sets (than in your test sets)
  • check quality of forecasted values on control sets
  • compute confidence intervals for individual errors (error defined e.g. as |true value minus forecasted value|) to make sure that error is small enough AND not too volatile (it has small variance across all control sets)

Related article:

Views: 33488

Comment

You need to be a member of AnalyticBridge to add comments!

Join AnalyticBridge

Comment by Frank Martins on September 8, 2013 at 6:42am

Vincent - you presented a nice, quick summary of the potential drawbacks of these techniques. However, you essentially claim that these techniques are used out of ignorance and due to a lag in awareness, which is often untrue, and may be a misleading statement. In some cases, some of these techniques are the best tool for the task.

For instance, in psychology, artificial neural networks are often utilized because they provide a rough analogy to biological neural networks. Their supposed "deficiencies" are actually useful traits in some cases - for example, overfitting can be taken advantage of for modeling experimental data because it sometimes matches what subgroups of participants tend to do on some tasks: you can go from modeling one group of study participants to another by tweaking parameters so as to encourage overfitting. The supposed instability of neural networks can be similarly used to one's advantage, or it can be mitigated by tweaking model parameters, or by averaging out a bunch of results.

Thus, the above article, while presenting some valid points, also presents a narrow perspective, and is consequently overly dismissive and misleading.

Comment by Ahmed Khamassi on January 20, 2013 at 10:23am
All the points are valid, but rather than saying the technique is rubbish I would say the process is. All techniques have shortcomings and are prone to be misused or to overfitting. The best way to go around this is to remind the user of the hypotheses of each technique and of the minimum best practice process: data discovery, transformation, variable selection and or reduction, modelling (preferably a few), validation, testing, control etc.
Comment by Vincent Granville on October 31, 2012 at 7:03pm

There's no miracle cure. My solution is to

  1. blend multiple models to identify as many significant  patterns as possible,  
  2. use multiple data sets including lists of events (with dates and event category) impacting business,
  3. use good cross-validation / model fitting / design of experiment,
  4. use proper metrics - both in your internal / external databases, as well as to measure lift and lift sustainability.
  5. Get good confidence intervals on anything that you measure. Keep in mind that if you have tons of confidence intervals, quite a few will provide false positives in the context of hypothesis testing.
Comment by Bill Luker Jr on October 31, 2012 at 6:39pm

So, Vincent, I respect your opinion, if only because you have so many that there must be some that are right, eh? Only joking, but just a little bit. (I am a very opinionated person too.) I echo the other commenter who asked you which approaches you favor. I am new to the data mining game (was taught, like many economists, that it was a no-no), so am particularly interested in what you might say.

Thanks

Bill Luker

Comment by Ralph Winters on October 3, 2012 at 12:48pm

Among these 8 worst techniques are the 5 BEST techniques (including linear regression).  Why?  They have stood the test of time and even today can be used to solve most of the worlds statistical problems. If you take the time to learn and master 3-4 of these techniques you are on your way to understanding the pitfalls that Vincent describes and can overcome them. What makes them the worst is not the techniques thamselves, but how they are employed.

-Ralph Winters

Comment by Bhagirath Addepalli on October 1, 2012 at 9:20pm

I am sure all the people discussing this subject are way more knowledgeable than me. I have only been in the academics, and am looking to make the transition to real-world data analysis. Given that, I would be interested if there could be article titled, "The 8 best predictive modelling techniques". My guess is that the answer to that would be: the goodness of a technique is relative to the problem that it is put to use on? There is no panacea for all ills? If that is true, shouldn't as much time be spent on understanding the problem at hand, as should be spent on scouting for techniques appropriate for the problem? But isn't it also true that in the industry, speed is sometime sacrificed for accuracy (which can only be achieved with any guarantee by better understanding the problem at hand) ? 

Comment by Tim Daciuk on October 1, 2012 at 5:59pm

Well Vincent, once again you've "set the cat amongst the pigeons"!

 

Great article.

Comment by Ramesh Hariharan on September 30, 2012 at 10:36pm

Vincent, thanks for the great post. Any tool in the hands of the semi-ignorant is a recipe for disaster, it doesn't matter whether you're working with BigData or traditional data. However, tree-based methods is one of my favorite techniques. Even though simple decision trees such as CART are quite unstable, there are other tree-based techniques that are pretty good for prediction, such as RandomForest. However, they're a bit of a black box, in that they may be great at prediction, but they are probably not very useful for explaining why a variable is important (except variable importance). We can do cross-validated decision trees that are more useful than plain simple CART. Moreover, instability of some of the techniques can be overcome using ensembles, but that's only for prediction, not for estimation.

Would like to see a follow-up post on the most suitable techniques for BigData.

Comment by K.Kalyanaraman on September 30, 2012 at 1:17am

Prediction and Statistical inference are two different issues. Robustness is concerned with parameter estimates through least square methods; the resulting sampling distributions of the estimators do not vary much even if linearity is violated to some extent. If prediction is considered, especially in the timeseries situation the lead time for prediction will be more important than other things. using data upto 2000 one may not predict for 2050.

Comment by Delyan Savchev on September 29, 2012 at 11:51am

We shouldn't blame the alphabet if there are illiterate people using it out there. We should educate the people. 

The 8 techniques that you have specified above have been separate research areas in statistics throughout the last 100 years. That is and still will be the basis for statistics education as most of the real life approximate ad-hoc actions a practicing statistician does are based on altering/combining these 8 paradigms. How are you going to use, say, generalized linear models if you don't know the linear ones.

In the real life case one should analyze the problem at hand, prior to selecting the method, but of course he should master the methods at the first place...

 

© 2014   AnalyticBridge.com is a subsidiary and dedicated channel of Data Science Central LLC

Badges  |  Report an Issue  |  Terms of Service